Abstract
For any autonomous driving vehicle, the control module determines its road performance and safety; its precision and stability must stay within a carefully designed range. Nonetheless, control algorithms require vehicle dynamics (such as longitudinal dynamics) as inputs, which are often obscure to calibrate in real time. Consequently, most research-oriented autonomous vehicles rely on manual, one-by-one calibrations to achieve reasonable performance. Since manual calibration is unsustainable for industrial mass production, we introduce a machine-learning-based auto-calibration system for autonomous driving vehicles. This paper details a data-driven longitudinal calibration procedure using machine learning techniques. We first generate offline calibration tables from human driving data, serving as an initial guess. We then employ an online-learning algorithm to update this table based on real-time performance analysis. This system has been deployed to over one hundred Baidu Apollo self-driving vehicles since April 2018, proving effective over thousands of hours and kilometers of testing.
Introduction
In autonomous driving, the longitudinal control module manages speed and braking. To operate accurately, the system must precisely map throttle or brake commands to the resulting acceleration at various speeds. Traditionally, this relationship is calibrated manually for every individual vehicle—a significant bottleneck for scaling production. Furthermore, vehicle dynamics are not static; factors such as cargo load, tire wear, and road surface friction cause a vehicle's behavior to drift over time. Static, manually-tuned tables cannot account for these variables, leading to increased "station error" (positional deviation) and "speed error."
System Architecture and Methodology
The Baidu Apollo auto-calibration system utilizes a two-stage hybrid approach to replace manual tuning with a scalable, data-driven pipeline.
1. Offline Model Generation
Rather than modeling complex physics (e.g., engine torque or transmission loss), the system treats the vehicle as a "black box." By collecting just 20 minutes of human driving data, a three-layer feedforward neural network is trained to map acceleration and speed to throttle/brake commands. This creates a robust initial calibration table that serves as a baseline for the fleet.
2. Online Learning and Adaptation
To handle real-time changes, such as a delivery vehicle carrying a heavy load, the system employs a model-free, gradient-descent-based algorithm. As the vehicle operates, it compares expected acceleration against actual sensor measurements, updating the calibration table on the fly. To ensure safety, the system enforces monotonicity constraints (ensuring more throttle always leads to more acceleration) and uses a similarity cost to prevent the table from deviating excessively from the human-verified baseline. This update cycle is highly efficient, taking less than 2 milliseconds.
Experimental Results
The system was validated across a diverse fleet, including Lincoln MKZ hybrid passenger cars and Neolix AX1 electronic delivery vehicles.
Accuracy and Superiority
The neural network achieved accuracy within 3% for the effective range of throttle and braking. In comparative road tests, the auto-calibration system significantly outperformed manual methods, reducing average station error from 0.588m to 0.362m.
Robustness to Load Changes
The system's adaptability was particularly evident during overload testing. When a Neolix AX1 was overloaded by 20% (360kg), the online learning module maintained a station error of 0.360m. Without online updates, the error ballooned to 0.778m, demonstrating the system's necessity for consistent performance in varying operational conditions.
Conclusion
By combining offline neural network initialization with real-time online learning, Baidu has developed a longitudinal calibration system capable of supporting industrial-scale autonomous fleets. This approach eliminates the need for tedious manual tuning while providing superior accuracy and adaptability to real-world mechanical wear and load variations. Future work will look toward extending these data-driven techniques to lateral (steering) calibration to further enhance vehicle stability and lane-keeping precision.
Summary updated